ESTRO 2024 - Abstract Book

S4565

Physics - Machine learning models and clinical applications

ESTRO 2024

1. Rossi L, Cambraia Lopes P, Marques et al. On the Importance of Individualized, Non-Coplanar Beam Configurations in Mediastinal Lymphoma Radiotherapy, Optimized With Automated Planning. Front Oncol. 2021 Apr 15;11:619929.

2798

Mini-Oral

Investigating the role of morphology in deep learning-based liposarcoma grading

Johannes Kiechle 1,2,3 , Sarah C. Foreman 4 , Stefan Fischer 1,2,5 , Daniel Rusche 1 , Verena Rösner 4 , Ann-Kathrin Lohse 6 , Carolin Mogler 7 , Stephanie E. Combs 1 , Marcus R. Makowski 4 , Klaus Woertler 4 , Daniel M. Lang 8 , Julia A. Schnabel 2,8,9 , Alexandra S. Gersing 10 , Jan C. Peeken 1,11,12 1 Klinikum Rechts der Isar, Department of Radiation Onkology, Munich, Germany. 2 Technical University of Munich, Institute for Computational Imaging and AI in Medicine, Munich, Germany. 3 Konrad Zuse School of Excellence in Reliable AI, Konrad Zuse School of Excellence in Reliable AI, Munich, Germany. 4 Klinikum Rechts der Isar, Department of Radiology, Munich, Germany. 5 Munich Center for Machine Learning, Munich Center for Machine Learning, Munich, Germany. 6 University Hospital Munich (LMU), Department of Radiology, Munich, Germany. 7 Klinikum Rechts der Isar, Institute of Pathology, Munich, Germany. 8 Helmholtz Munich, Institute of Machine Learning in Biomedical Imaging, Munich, Germany. 9 King's College London, School of Biomedical Engineering and Imaging Sciences, London, United Kingdom. 10 University Hospital Munich (LMU), Department of Diagnostic and Interventional Neuroradiology, Munich, Germany. 11 Helmholtz Munich, Institute of Radiation Medicine, Munich, Germany. 12 Deutsches Konsortium für Translationale Krebsforschung (DKTK), Partner Site Munich, Heidelberg, Germany Soft-tissue tumors encompass a heterogeneous group of tumors with a large variety of subtypes, involving benign entities to locally aggressive counterparts known as soft-tissue sarcomas (STS). The latter are prone to metastasize and can implicate a significant reduction in survival rate. For patients diagnosed with STS, tumor subtype identification, typically achieved through pathological examination following focal biopsies, constitutes a decisive factor in shaping the optimal follow-up clinical treatment strategy. The decision to use radiation therapy for STS is based on the tumor subtype, location, stage, and the potential for complete surgical removal. For high-grade STS such as G2 and G3 liposarcoma and cases where complete resection is not feasible or uncertain, radiation therapy plays a crucial role. It can be administered preoperatively or post operatively, forming an integral part of the comprehensive treatment strategy. Conversely, for atypical lipomatous tumors (ALT), also known as well-differentiated liposarcomas, surgical intervention represents the primary therapeutic approach. In contrast to conventional focal biopsies, artificial intelligence-based methods applied to medical imaging applications pose an alternative way to characterize STS. Previous research has underscored the relevance of texture radiomics features or convolutional neural networks for STS grading [1]. In this work, however, we aimed to investigate the potential of differentiating ALTs from high-grade liposarcomas solely based on their 3D morphological characteristics. Therefore, we conducted a comparative analysis between a machine learning (ML) approach utilizing radiomics shape features and a deep learning (DL) approach employing graph convolution neural networks for tumor subtype classification into ALTs and high-grade liposarcomas. Purpose/Objective: